1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 2);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc1x0123 = vacc0x0123;
52 int32x4_t vacc1x4567 = vacc0x4567;
53
54 size_t k = kc;
55
56
57 while (k >= 8 * sizeof(int8_t)) {
58 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
59 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
60
61 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69
70 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
71 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
72
73 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
74 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
75 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
76 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
77 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
78 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
79 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
80 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
81 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
82 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
83
84 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
85 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
86 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
87 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
88 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
89 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
90 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
91 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
92 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
93 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
94
95 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
96 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
97 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
98 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
99 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
100 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
101 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
102 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
103 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
104 const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
105
106 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
107 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
108 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
109 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
110 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
111 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
112 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
113 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
114
115 k -= 8 * sizeof(int8_t);
116 }
117
118 if XNN_UNLIKELY(k != 0) {
119 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
120 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
121
122 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124
125 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
126 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
127 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
128 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
129 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
130 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
131 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
132 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
133 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
134 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
135
136 if (k > 2 * sizeof(int8_t)) {
137 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139
140 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
141 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
142 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
143 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
145 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
146 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
147 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
148 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
149 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
150
151 if (k > 4 * sizeof(int8_t)) {
152 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154
155 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
156 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
157 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
158 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
159 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
160 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
161 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
162 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
163 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
164 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
165 }
166 }
167 }
168
169 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
170 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
171 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
172
173 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
174 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
175 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
176 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
177
178 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
179 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
180 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
181 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
182
183 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
184 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
185 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
186 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
187
188 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
189 #if XNN_ARCH_ARM64
190 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
191 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
192
193 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
194 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
195
196 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
197 #else
198 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
199 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
200
201 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
202 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
203
204 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
205 #endif
206
207 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
208 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
209
210 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
211 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
212
213 if (nc >= 8) {
214 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
215 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
216
217 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
218 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
219
220 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
221 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
222
223 nc -= 8;
224 } else {
225 // Final case where not all of the 8 columns fit in the destination.
226 if (nc & 4) {
227 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
228 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
229 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
230 }
231 if (nc & 2) {
232 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
233 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
234 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
235 }
236 if (nc & 1) {
237 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
238 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
239 }
240
241 nc = 0;
242 }
243 } while (nc != 0);
244 }
245